Method and Apparatus for Processing Chip Based on Deep Learning

1. A method for processing a chip based on deep learning, comprising steps: 1) scanning the chip with femtosecond laser in a predetermined polarization state to produce a main scanning trajectory and periodic nano-stripes on both sides of the main scanning trajectory, so as to form a nano-ridge structure on a surface of the chip; 2) obtaining a super-resolution microscopic image of the nano-ridge structure by super-resolution microscopy, removing an invalid area around the microscopic image, retaining a periodic nano-stripe area in a center of the microscopic image, and scaling the microscopic image to a preset size to obtain a target image; 3) reconstructing the target image based on deep learning for image super-resolution to obtain the reconstructed image, and recognizing and processing the reconstructed image to obtain characteristic parameters of the nano-ridge structure as input parameters (x1, x2, . . . , xi) of an input layer X of a deep learning model for femtosecond laser processing; 4) directing the input parameters of the input layer X into a neutral network including M hidden layers h1, h2, . . . , hm for deep learning for femtosecond laser processing, activating a respective hidden layer by performing nonlinear transformation on input parameters of a previous layer with a nonlinear activation function f(W, b), where W represents a weight value, and b represents an offset value, and performing deep learning in the M hidden layers to obtain output values (y1, y2, . . . , yj) of an output layer Y of the deep learning model for femtosecond laser processing; 5) evaluating inconsistency between the output values (y1, y2, . . . , yj) of the output layer Y of the deep learning model for femtosecond laser processing and a target value O using a loss function L to obtain a loss function value θ represented by a formula: θ=L(Y,O) so as to complete a training process for deep learning; 6) adjusting corresponding processing parameters of the chip according to the output values (y1, y2, . . . , yj) of the output layer Y, optimizing the weight value W and the offset value b of the deep learning model, and repeating step 2) to step 5) for repeated training for femtosecond laser processing to minimize the loss function value θ; and 7) outputting the optimized nano-ridge structure.

2. The method of claim 1, wherein the femtosecond laser has a pulse width equal to or less than 200 fs, a repetition frequency of 1 KHz to 1 GHz, and a laser energy of 0 to 100 mW.

3. The method of claim 1, wherein the predetermined polarization state is a predetermined angle between a laser polarization direction and a scanning direction, and the predetermined angle is kept unchanged by the deep learning for femtosecond laser processing during the scanning.

4. The method of claim 3, wherein the chip is scanned by: keeping the position of the chip unchanged, controlling laser to scan the chip along the main scanning trajectory with a scanning direction varied with the scanning trajectory, and keeping an angle between the laser polarization direction and the scanning direction unchanged; keeping the focus position of laser unchanged, controlling the chip to move along the main scanning trajectory, and adjusting the position and direction of the chip so as to keep an instant movement direction of the chip with respect to laser constant; or keeping the scanning direction of laser unchanged, adjusting the position and direction of the chip, and forming the main scanning trajectory while keeping the scanning direction unchanged.

5. The method of claim 1, wherein the scanning speed ranges from 100 to 1000 μm/s.

6. The method of claim 1, wherein the periodic nano-stripes of the nano-ridge structure are equally spaced, and a spacing between two adjacent nano-stripes ranges from 10 nm to 1 μm.

7. The method of claim 6, wherein the spacing between two adjacent nano-stripes ranges from 10 to 200 nm.

8. The method of claim 1, wherein an angle between a main scanning trajectory and each of the periodic nano-stripes ranges from 0 to 90°.

9. The method of claim 8, wherein the angle between the main scanning trajectory and each of the periodic nano-stripes is 90°.

10. The method of claim 1, wherein the super-resolution microscopy is super-lens imaging based on surface plasmon resonance, stimulated emission depletion microscopy (STED), photoactivated localization microscopy (PALM), structured illumination microscopy (SIM), or stochastic optical reconstruction microscopy (STORM).

11. The method of claim 1, wherein the super-resolution microscopic image of the nano-ridge structure is obtained by the deep learning for image super-resolution based on a deep learning model for image super-resolution, and the processing parameters of the nano-ridge structure are adjusted and optimized by the deep learning for femtosecond laser processing based on a deep learning model for femtosecond laser processing, which is initialized and optimized during the femtosecond laser processing.

12. The method of claim 1, wherein the periodic nano-stripe of the nano-ridge structure in the reconstructed image is visible in a predetermined image size; and recognizing and processing the reconstructed image comprises: recognizing the reconstructed image, performing data enhancement, and performing data preprocessing to obtain one-dimensional vectors as the input parameters (x1, x2, . . . , xi), wherein the input parameters correspond to the characteristic parameters of the nano-ridge structure comprising a length, a spacing, and a parallelism degree of the periodic nano-stripes.

13. The method of claim 1, wherein activating the respective hidden layer by performing nonlinear transformation on input parameters of the previous layer with the nonlinear activation function f(W, b) comprises: activating the hidden layers h1, h2, . . . , hm with the corresponding nonlinear activation functions:, h 1 = f ⁡ ( W 1 ⁢ X + b 1 ) ⁢ ⁢ h 2 = f ⁡ ( W 2 ⁢ h 1 + b 2 ) … h m = f ⁡ ( W m ⁢ h m - 1 + b m ) where W1, W2, . . . , Wm represent weight values of the respective hidden layers, and b1, b2, . . . , bm represent offset values of the respective hidden layers, performing deep learning in the M hidden layers comprises: performing deep learning in the hidden layers h1, h2, . . . , hm to obtain one-dimensional vectors as the output values (y1, y2, . . . , yj) of the output layer Y: Y=f(Wm+1hm+bm+1) where the one-dimensional vectors (y1, y2, . . . , yj) correspond to processing parameters comprising a laser energy, a repetition frequency, a polarization state, and a scanning speed, and Wm+1 and bm+1 represent a weight value and an offset value of the output layer, respectively.

14. The method of claim 1, wherein the super-resolution microscopic image is taken at a frequency greater than or equal to 50 fps.

15. An apparatus for processing a chip based on deep learning, comprising: a processing platform, on which a chip to be processed is placed and fixed; a laser device configured to emit femtosecond laser with adjustable pulse width, laser energy, repetition frequency and polarization state; a scanning device movable in a direction in an XY plane, and configured to position a laser spot and scan the chip with the laser spot; a focusing device configured to focus the femtosecond laser; an imaging device configured to obtain a super-resolution microscopic image of a nano-ridge structure on a surface of the chip, and obtain a target image by removing an invalid area around the microscopic image, retaining a periodic nano-stripe area in a center of the microscopic image, and scaling the microscopic image to a preset size; an adjusting device configured to adjust a translation of the processing platform in an X, Y or Z direction, or a rotation of the processing platform about a Z axis; a positioning device configured to position the chip; and a controller configured to control the laser device, the scanning device, the imaging device, the adjusting device, and the positioning device, and comprising a deep learning unit embedded therein, wherein the deep learning unit is configured to: reconstruct the target image obtained by the imaging device based on deep learning for image super-resolution to obtain the reconstructed image, and recognize and process the reconstructed image to obtain characteristic parameters of the nano-ridge structure as input parameters (x1, x2, . . . , xi) of an input layer X of a deep learning model for femtosecond laser processing; direct the input parameters of the input layer X into a neutral network including M hidden layers h1, h2, . . . , hm for deep learning for femtosecond laser processing, activate a respective hidden layer by performing nonlinear transformation on input parameters of a previous layer with a nonlinear activation function f(W, b), where W represents a weight value, and b represents an offset value, and perform deep learning in the M hidden layers to obtain output values (y1, y2, . . . , yj) of an output layer Y of the deep learning model for femtosecond laser processing; and evaluate inconsistency between the output values (y1, y2, . . . , yj) of the output layer Y of the deep learning model for femtosecond laser processing and a target value O using a loss function L to obtain a loss function value θ represented by a formula: θ=L(Y,O) so as to complete a training process for deep learning; and adjust corresponding processing parameters of the chip according to the output values (y1, y2, . . . yj) of the output layer Y, optimize the weight value W and the offset value b of the deep learning model, and perform repeated training to minimize the loss function value θ.

16. The apparatus of claim 15, wherein the imaging device is a super-resolution microscope, and the super-resolution microscopic image is taken at a frequency greater than or equal to 50 fps.

17. The apparatus of claim 15, wherein the chip is scanned by: adjusting a scanning speed and a scanning direction by the scanning device; adjusting the scanning speed and the scanning direction by the adjusting device; adjusting the scanning speed by the scanning device, and adjusting the scanning direction by the adjusting device; or a combination thereof.

18. The apparatus of claim 15, wherein the femtosecond laser has a pulse width equal to or less than 200 fs, a repetition frequency of 1 KHz to 1 GHz, and a laser energy of 0 to 100 mW.

19. The apparatus of claim 15, wherein the scanning speed ranges from 100 to 1000 μm/s.

20. The apparatus of claim 15, wherein the periodic nano-stripes of the nano-ridge structure are equally spaced, and a spacing between two adjacent nano-stripes ranges from 10 to 200 nm.